Deploy Cloud Harness MCP to a VPS, gate the hostname with Cloudflare Access, turn on Managed OAuth, then wire up the ChatGPT connector. With 21 real screenshots, identifying data blurred.
The Advisor is a supervisory control loop running alongside the main agent in the same session, not a stronger model called in to give an opinion.
A practical guide to using Archify with an AI agent for code-grounded diagrams or proposed flows, using one bakery example throughout.
Install, validate, choose modes, write prompts, and operate UI Craft end to end.
OMP /plan locks the runtime read-only, AgentKit ak:plan produces a durable file-backed roadmap. When to use which, and why not to nest the two workflows.
OMP roles are routing aliases, not agents. Assign models by workload, inspect effective state, and retry through fallbacks.
Install the imagegen skill, pick one of seven visual directions, and generate infographics that stay on content.
Three independent commands, pi-dev, pi-ak, and pi-omp, with user state split by Mise.
Agent-aware terminal multiplexer: keep sessions on your machine and attach from desktop, SSH, or phone.
Do not copy the kit into every project. The manifest picks a channel; the control plane keeps immutable releases.
One entry point routes every cmux request. Identity before mutation. Additive layout with --focus false.
Run several Codex threads in parallel, each in its own worktree, with paste-ready prompts.
Import, adapt, link, and retire AI skills across runtimes without losing provenance.
preview orchestrates thinking, frontend-design executes UI, tech-graph and show-off handle publishing. Includes a real-run case study.
When an agent can read files, run shell commands, edit code and ship PRs, one line of instruction is not enough. You need a check layer before the tool runs.
From AI assistant to development partner. How Claude works, and how to use the Engineer Kit through real-world use cases.
The distance between prompt and codebase is unknowns. A six-stage pipeline, three fixed gates around ck:cook, and the cost of choosing a model per stage.
--deep helps the plan see the full scope, --tdd adds a regression gate for refactors. When each option pays off, and when you can skip them.